Research On Intelligent Quantitative Trading Strategies

Portfolio management(PM)refers to the process of effectively managing large-scale capital through investment strategies and portfolios,with the ultimate goal of achieving significant investment returns for clients and investors.Asset management strategies are mainly divided into two categories: active investment and passive investment.Each strategy includes two key tasks: asset selection and asset allocation.The purpose of the asset selection task is to pick suitable assets for investment,while the asset allocation task focuses on achieving optimal capital distribution among selected assets.In active investment strategies,investors or fund managers actively seek returns that surpass the market average,thus the optimization objectives for asset selection and allocation aim to maximize investment returns through precise market analysis and asset evaluation.Unlike maximizing investment returns,the goal of passive investment strategies is to achieve stable growth in line with the overall market.In passive investment strategies,index investing is particularly favored,aiming to precisely replicate the performance of market indices through partial index tracking(PIT).Due to the differences in investment purposes from active investment strategies,in passive investments,the optimization objectives for asset selection and allocation turn towards how to accurately track market trends and performance.This dissertation aims to explore how advanced artificial intelligent technologies,such as graph neural network(GNN)and deep reinforcement learning(DRL),can be integrated into active and passive investment strategies to optimize the two core tasks of asset selection and allocation.This dissertation not only explores the challenges and solutions in their practical applications but also analyzes the potential of these technologies in enhancing the efficiency and accuracy of asset management.Specifically,the main research contents of this dissertation include the following points:(1)Active investment–asset selection based on relation temporal graph convolutional network: For the asset selection task in active investment,this dissertation designs a relation temporal graph convolutional network(RT-GCN)to overcome two major limitations of traditional asset selection methods:(i)treating each asset as an independent entity,thereby neglecting the interrelationships between assets,and(ii)modeling asset prediction as either a regression task(predicting asset prices)or a classification task(predicting asset trends)rather than selecting the assets with the highest expected return from the market.Firstly,we model the relationships among assets and their temporal features into a relation-temporal graph.Then,we apply RT-GCN to extract relation-temporal features for each asset.Finally,based on the extracted features,each asset is scored and ranked,with higher-scoring assets indicating higher future investment returns.Investors can then invest in top-scoring assets as needed to maximize returns.(2)Active investment–asset allocation by integrating diverse investment horizon: For the asset allocation task in active investment,this dissertation introduces a trading approach integrated diverse horizons to relieve the non-stationarity issues.The inherent non-stationarity of financial markets primarily arises from the diversity in traders’ investment horizons and strategies,a factor not adequately considered in existing machine learning-based solutions.Therefore,this dissertation employs reinforcement learning,integrating multiple investment insights into the trading strategy to adapt to market changes.Initially,it tailor information specific to certain horizons to learn multiple sub-policies.This allows each sub-policy to recognize patterns within its respective horizon and make insightful pre-decisions.Subsequently,we develop a cross-horizon strategy that makes final trade decisions by integrating pre-decisions from the specific horizon policies learned in the first step.By incorporating pre-decisions from diverse investment horizons into the decision-making process,this strategy enhances its adaptability to market changes.(3)Passive investment–asset selection based on reinforcement learning: For the asset selection task in PIT,this dissertation designs a solution based on deep reinforcement learning to select assets,addressing the inefficiencies of traditional heuristic methods,especially in large-scale index tracking.This study models the asset selection problem as a markov decision process(MDP)and optimizes it using deep reinforcement learning(DRL).The dissertation also proposes a relevance-aware deep Q-network(RA-DQN)and a cost-aware reward function.RA-DQN intelligently selects a specified number of assets using domain knowledge,while the cost-aware reward function evaluates the selections and guides RA-DQN to minimize tracking error and control transaction costs.(4)Passive investment–index tracking based on hierarchical reinforcement learning: For the asset selection and allocation tasks in PIT,this dissertation introduces a hierarchical model for PIT(named HIT)to ensure the collaborative optimization between asset selection and allocation.In this study,PIT is formulated as a hierarchical markov decision process,optimized through hierarchical reinforcement learning(HRL).The model comprises a high-level policy that learns to select assets from constituents,and a lowlevel policy that handles the task of asset allocation among the selected assets.The model is equipped with a cost-sensitive reward function,serving as a link for the collaborative optimization of both policies.By learning separate strategies for the two tasks and utilizing the reward function as a bridge for collaborative optimization,the model ensures collaborative optimization of two tasks.Finally,this dissertation summarizes the research content and provides a future outlook based on feasible research ideas and directions.